This PR enhances the dataset card for MMCricBench by:
- Adding a direct link to the research paper, [Mind the (Language) Gap: Towards Probing Numerical and Cross-Lingual Limits of LVLMs](https://huggingface.co/papers/2508.17334), in the introductory section.
- Including the `image-text-to-text` task category in the metadata, providing a more general classification for this VQA dataset.
- Removing the redundant link to the dataset's own Hugging Face page, which is implicitly the current page.